List your company
ResearchFormerly WizWand

Image Classification on ImageNet-1K (Fine-tuning)

86.8Accuracy (FT)Softmax
55.663.771.879.9May 27, 2022Jan 13, 2023Sep 2, 2023Apr 21, 2024Dec 8, 2024Jul 28, 2025Mar 17, 2026
Updated 7mo ago

Evaluation results

MethodLinks
SoftmaxBackbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=310.60, Peak Mem. (GB)=1.3181, Throughput (imgs/s)=36.52, fine-tuning=true2026.0386.8
ViT-AdaLA (Ours)Backbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=262.19↓15.6%, Peak Mem. (GB)=1.2163↓7.7%, Throughput (imgs/s)=41.56↑16.1%, fine-tuning=true2026.0385.5
MILANBackbone=ViT-B, Supervision=CLIP-B, Forward Ratio=25%, Reconstruct Ratio=100%, Epochs=4002022.1185.4
MaskAlignBackbone=ViT-B, Supervision=CLIP-B, Forward Ratio=30%, Reconstruct Ratio=0%, Epochs=2002022.1185.4
CMAEBackbone=CVIT-B, Supervision=RGB, Forward Ratio=25%, Reconstruct Ratio=75%, Epochs=16002022.1185.3
MixMIMBackbone=MixMIM-B, Supervision=RGB, Forward Ratio=100%, Reconstruct Ratio=100%, Epochs=6002022.1185.1
MCMAEBackbone=CVIT-B, Supervision=RGB, Forward Ratio=25%, Reconstruct Ratio=75%, Epochs=16002022.1185
BEIT V2Backbone=ViT-B, Supervision=CLIP-B, Forward Ratio=100%, Reconstruct Ratio=40%, Epochs=3002022.1185
MaskDistillBackbone=ViT-B, Supervision=CLIP-B, Forward Ratio=100%, Reconstruct Ratio=40%, Epochs=3002022.1185
FD-CLIPBackbone=ViT-B, Supervision=CLIP-B, Forward Ratio=100%, Reconstruct Ratio=0%, Epochs=3002022.1184.9
MVPBackbone=ViT-B, Supervision=CLIP-B, Forward Ratio=100%, Reconstruct Ratio=40%, Epochs=3002022.1184.4
C2FMAEModel=ViT-B, Modality=RGB/Inst./Sem., Masking=Progressive, PT Epoch=1600, PT Cost=∼5.2×, Fine-tuning Input=multi-granular data, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0384.4
data2vecBackbone=ViT-B, Supervision=EMA, Forward Ratio=100%, Reconstruct Ratio=60%, Epochs=8002022.1184.2
C2FMAEModel=ViT-B, Modality=RGB/Inst./Sem., Masking=Progressive, PT Epoch=1600, PT Cost=∼5.2×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0384.2
MIRLModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=800, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0384.1
SimMIMBackbone=Swin-B, Supervision=RGB, Forward Ratio=100%, Reconstruct Ratio=60%, Epochs=8002022.1184
DMAEBackbone=ViT-B, Supervision=MAE-L, Forward Ratio=25%, Reconstruct Ratio=75%, Epochs=1002022.1184
iBOTModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=1600, PT Cost=∼5.7×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0384
MaskFeatModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=1600, PT Cost=∼20.1×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0384
ROPIMModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=800, PT Cost=∼10.4×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0384
ConMIMModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=800, PT Cost=∼4.4×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.7
C2FMAEModel=ViT-B, Modality=RGB/Inst./Sem., Masking=Progressive, PT Epoch=400, PT Cost=∼1.3×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.7
MAEBackbone=ViT-B, Supervision=RGB, Forward Ratio=25%, Reconstruct Ratio=75%, Epochs=16002022.1183.6
CAEBackbone=ViT-B, Supervision=DALLE, Forward Ratio=25%, Reconstruct Ratio=75%, Epochs=8002022.1183.6
MaskFeatBackbone=ViT-B, Supervision=HOG, Forward Ratio=100%, Reconstruct Ratio=40%, Epochs=3002022.1183.6
MAEModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=1600, PT Cost=∼4.0×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.6
CAEModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=800, PT Cost=∼4.6×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.6
ViT-AdaLA (Stage 2)Backbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=262.19↓15.6%, Peak Mem. (GB)=1.2163↓7.7%, Throughput (imgs/s)=41.56↑16.1%, fine-tuning=true2026.0383.4
SemMAEModel=ViT-B, Modality=RGB, Masking=Semantic, PT Epoch=800, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.3
AutoMAEModel=ViT-B, Modality=RGB, Masking=Semantic, PT Epoch=800, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.3
MultiMAEModel=ViT-B, Modality=RGB/Dep./Sem., Masking=Random, PT Epoch=1600, PT Cost=∼5.2×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.3
BEITBackbone=ViT-B, Supervision=DALLE, Forward Ratio=100%, Reconstruct Ratio=40%, Epochs=8002022.1183.2
MoCo v3Model=ViT-B, Modality=RGB, PT Epoch=300, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.2
BEiTModel=ViT-B, Modality=RGB, Masking=Random, PT Epoch=800, PT Cost=∼7.0×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.2
MFMModel=ViT-B, Modality=RGB/Frequency, Masking=Random, PT Epoch=300, PT Cost=∼1.1×, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0383.1
MAE (reproduced)Model=ViT-B, Modality=RGB, Masking=Random, PT Epoch=400, PT Cost=∼1.0×, Note=reproduced using official code, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0382.9
UnMAEModel=ViT-B, Modality=RGB, Masking=Uniform, PT Epoch=200, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0382.9
DINOModel=ViT-B, Modality=RGB, PT Epoch=300, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0382.8
MultiMAE (reproduced)Model=ViT-B, Modality=RGB/Dep./Sem., Masking=Random, PT Epoch=400, PT Cost=∼1.3×, Note=reproduced using official code, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0382.7
ScratchModel=ViT-B, Input Size=224 × 224, Evaluation Protocol=Fine-tuning2026.0382.3
MonarchBackbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=269.44↓13.3%, Peak Mem. (GB)=1.2304↓6.7%, Throughput (imgs/s)=18.31↓48.9%, fine-tuning=true2026.0381.5
NystromformerBackbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=265.20↓14.6%, Peak Mem. (GB)=1.2163↓7.7%, Throughput (imgs/s)=18.29↓50.8%, fine-tuning=true2026.0380.7
CIMTarget=BEIT, Epochs PT=300, Backbone=ResNet-50, Fine-tuning=true2022.0580.4
A2MIMTarget=RGB, Epochs PT=300, Backbone=ResNet-50, Fine-tuning=true2022.0580.4
SwAVTarget=CL, Epochs PT=800, Backbone=ResNet-50, Fine-tuning=true2022.0580.2
BYOLTarget=CL, Epochs PT=400, Backbone=ResNet-50, Fine-tuning=true2022.0580.1
SimSiamTarget=CL, Epochs PT=400, Backbone=ResNet-50, Fine-tuning=true2022.0580
SimCLRTarget=CL, Epochs PT=800, Backbone=ResNet-50, Fine-tuning=true2022.0579.9
Barlow TwinsTarget=CL, Epochs PT=800, Backbone=ResNet-50, Fine-tuning=true2022.0579.9
SimMIM+Target=RGB, Epochs PT=300, Backbone=ResNet-50, Fine-tuning=true, Reproduced=true2022.0579.9
Sup.Target=Label, Epochs PT=90, Backbone=ResNet-50, Fine-tuning=true2022.0579.8
MoCoV2Target=CL, Epochs PT=800, Backbone=ResNet-50, Fine-tuning=true2022.0579.8
CosformerBackbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=265.41↓14.6%, Peak Mem. (GB)=1.2226↓7.2%, Throughput (imgs/s)=39.55↑8.2%, fine-tuning=true2026.0370.6
PerformerBackbone=CLIP-L, Res.=512², Params (M)=304.20, FLOPS (G)=265.41↓14.6%, Peak Mem. (GB)=1.2164↓7.7%, Throughput (imgs/s)=36.48↑0.1%, fine-tuning=true2026.0369.9
HedgehogBackbone=CLIP-L, Res.=512², Params (M)=305.77, FLOPS (G)=265.41↓14.5%, Peak Mem. (GB)=1.2221↓7.3%, Throughput (imgs/s)=37.44↑2.5%, fine-tuning=true2026.0367.4
LoLCATSBackbone=CLIP-L, Res.=512², Params (M)=305.77, FLOPS (G)=265.41↓14.5%, Peak Mem. (GB)=1.2221↓7.3%, Throughput (imgs/s)=37.44↑2.5%, fine-tuning=true2026.0367.21
LinformerBackbone=CLIP-L, Res.=512², Params (M)=305.77, FLOPS (G)=262.19↓15.6%, Peak Mem. (GB)=1.2221↓7.3%, Throughput (imgs/s)=45.41↑25.3%, fine-tuning=true2026.0356.8